AI Techniques for Digital Twins: Methods and Applications in Complex Systems

Digital Twins (DTs) have emerged as a powerful paradigm for modeling, simulation, and decision-making in complex real-world systems by tightly coupling physical assets with their virtual counterparts. However, the development of effective DTs remains challenging due to the intrinsic complexity of physical processes, the presence of uncertainty, limited observability, and the need to balance accuracy, computational efficiency, and interpretability. This thesis investigates how Artificial Intelligence (AI) techniques can be systematically integrated into DT frameworks to address these challenges across heterogeneous application domains. The work focuses on four complementary research problems that exemplify key roles of AI within DT architectures. First, a Deep Learning (DL)-based super-resolution framework is proposed to enhance the spatial resolution of operational oceanographic models in coastal regions. By reconstructing fine-scale coastal dynamics from coarse-resolution inputs, the method improves the effective usability of existing products without requiring additional high-resolution numerical simulations. Second, the thesis addresses the modeling of robotic locomotion on deformable terrains. A data-driven surrogate modeling approach is developed to approximate the dynamics of wheel–soil interaction using real-world measurements. This solution provides a computationally efficient alternative to expensive physics-based models while preserving sufficient accuracy for simulation. Third, a framework for safety-critical scenario generation in autonomous driving is introduced. The proposed method combines diffusion-based generative models with formal spatio-temporal logic specifications to generate realistic yet targeted multi-agent traffic scenarios. By explicitly encoding safety requirements and domain knowledge through formal logic, the framework enables systematic and interpretable stress testing of autonomous driving systems, addressing the statistical inefficiency of traditional validation strategies. Lastly, we address the problem of Symbolic Regression, introducing cellular-inspired extensions of different semantic Genetic Programming (GP) variants to enhance diversity and reduce premature convergence in populations. An extensive experimental evaluation systematically analyzes the impact of different design choices on the trade-offs between predictive accuracy and model complexity, used as a proxy for interpretability, providing actionable insights for practitioners in real-world applications. Taken together, the contributions of this thesis show that AI techniques can be systematically integrated to tackle the core challenges of DTs. The results highlight the value of application-driven methodological innovation and provide practical solutions for DT deployment in various domains.